Publication Details

IEEE Access

Contribution To Journal


In medical imaging, automated landmark detection estimates the position of anatomical points in images to derive measurements. Previous approaches commonly employ coordinate regression. Landmark segmentation, a technique in which masks centered at the target point are segmented, has recently shown promising results. Here, we present segmentation-guided coordinate regression, a methodology that fuses both approaches and balances accuracy and robustness. Our approach identifies masks centered at landmarks using a segmentation network. Then, a coordinate regression network estimates the coordinates by employing the input image and the segmentation output. We assessed the methodology’s performance by detecting eight landmarks in full lower limb X-rays and investigated the impact of weight initialization, network backbone, and optimization of the loss function. The approach was contrasted with landmark segmentation and coordinate regression and applied to the analysis of lower limb malalignment. Results showed that deeper pretrained models with a weight of 0.2 at the segmentation loss detected landmarks more accurately. Segmentation-guided regression outperformed coordinate regression. Landmark segmentation was hampered by undetected landmarks and false positives. Due to its architecture, the proposed method did not suffer from failed detections, allowing lower limb malalignment to be reliably calculated. With respect to comparable literature, our approach leads to similar or improved results for landmark detection, translating to highly accurate and reliable lower limb malalignment analysis. In conclusion, we proposed a novel method for detecting landmarks in X-rays, which leads to a balance in accuracy and robustness and allows the measurement of lower limb malalignment.

DOI scopus